To increase our understanding of how humans have altered the Earth's surface and to facilitate land surface modeling experiments aimed to elucidate the direct impact of land cover change on the Earth system, we create and analyze a database of global land use/cover change (LUCC). From a combination of sources including satellite imagery and other remote sensing, ecological modeling, and country surveys, we adapt and synthesize existing maps of potential land cover and layers of the major anthropogenic land covers, including a layer of wetland loss, that are then tailored for land surface modeling studies. Our map database shows that anthropogenic land cover totals to approximately 40% of the Earth's surface, consistent with literature estimates. Almost all (92%) of the natural grassland on the Earth has been converted to human use, mostly grazing land, and the natural temperate savanna with mixed C3/C4 is almost completely lost (∼90%), due mostly to conversion to cropland. Yet the resultant change in functioning, in terms of plant functional types, of the Earth system from land cover change is dominated by a loss of tree cover. Finally, we identify need for standardization of percent bare soil for global land covers and for a global map of tree plantations. Estimates of land cover change are inherently uncertain, and these uncertainties propagate into modeling studies of the impact of land cover change on the Earth system; to begin to address this problem, modelers need to document fully areas of land cover change used in their studies.
 As a result of continuing population and economic growth, humans are altering huge swaths of the Earth's land surface through conversion of natural landscapes to cropland, built-up land, grazing land, inundated land, reservoirs, and tree plantations. It has been estimated that humans alter 39–50% of the Earth's land surface [Vitousek et al., 1997]. Land cover is dominated by grazing, which takes place on roughly one fourth of the terrestrial Earth.
 Changes to the land surface impact major biogeochemical cycles [Holmes et al., 2005], and directly alter climate through alterations to land-atmosphere fluxes of the water, energy, and carbon cycles [Bosch and Hewlett, 1982; Bounoua et al., 2002; Foley et al., 2005; Sterling, 2005]; therefore, a key research direction is to improve our knowledge on how global anthropogenic land use/cover change (LUCC) impacts our planet [GWSP, 2005]. To reach this goal, we must understand the nature of global LUCC; most basically, the actual areas altered, spatial distribution of these areas, and the types of LUCC. The most common tool used in global-scale studies on the impact of LUCC on the Earth system are land surface models (LSMs), a set of equations which simulate global-scale land surface water and energy fluxes that may be coupled to an atmospheric general circulation model, or driven by an atmospheric forcing data set.
 One cause of the areas of altered land being too low in LSM experiments is that the suite of major land cover changes generally is not considered in modeling efforts. For example, many studies only consider one or two types of anthropogenic land cover (ALC), here defined as land use/land covers of anthropogenic origin, including cropland, irrigated cropland, grazing land, and reservoirs. Cropland is often only considered without separate consideration of grazing or built-up land [Betts, 2001; Brovkin et al., 1999; Chase et al., 2000; DeFries, 2002; DeFries et al., 2002; Govindasamy et al., 2001; Matthews et al., 2003]. In addition to the estimation of appropriate areas of LUCC, it is important to include the range of major ALCs, since different types of LUCC alter different ecosystem and land surface processes and even the direction of change of surface fluxes. For example, conversion of savanna to grazing land should decrease evapotranspiration, while conversion of savanna to inundated land in reservoirs would increase evapotranspiration [Sterling, 2005]. Similarly, deforestation to irrigated cropland would invoke different climatological and ecological changes than would deforestation to built-up areas.
 While hydrologic modifications to the land surface are of potentially large importance to the energy and water cycles, they are only recently beginning to be considered in LSM studies, for example, irrigation and reservoir creation [Haddeland et al., 2006; Hanasaki et al., 2006]. And while there have been some studies on wetland loss since the last glacial maximum [de Noblet-Ducoudré et al., 2002], to our knowledge, studies on the impact of wetland loss driven by the current extent of land cover change have not yet been conducted.
 Another issue in mapping LUCC for LSMs is that the land cover classes in maps need to be coherent with ecological functioning categories used by LSMs. Plant functional types (PFTs), plant species or land covers sharing similar properties as regards their structure, photosynthesis pathway, response to disturbances and interactions with the land surface [Crucifix et al., 2005], are used as ecological functioning categories by many third generation land surface models [Bonan et al., 2002; Ducoudré et al., 1993a; Foley et al., 1996; Haxeltine and Prentice, 1996; Kucharik et al., 2000; Neilson, 1995; Running and Coughlan, 1988; Running and Gower, 1991; Running and Hunt, 1993; Schimel and Braswell, 1997; VEMAP Members, 1995] which translate land cover classes provided in maps to combinations of PFTs, in order to predict the composition and function of ecosystems. Land cover data sets do not come with rules to translate to PFTs, and such translation can add uncertainty and bias as the coherence between land cover types and PFTs is often unclear. For example, with the widely used data sets of Global Ecosystems Legend [Loveland et al., 2000; Wilson and Henderson-Sellers, 1985], many land cover classes are composed of mixed life-form classes (for example, “cool fields and woods,” “evergreen broadleaf cropland,” “pasture + tree”), making it difficult to choose percentage composition for the PFTs. This lack of coherence in PFTs and land covers results in ad hoc correlations and instances where each modeler creates his/her own rules, adding to the difficulty in comparing the results from different studies. Recent work has been done to document the translation among land cover classes to PFTs for the Canadian Terrestrial Ecosystem (CTEM) land surface model for time series of land cover maps from 1850 and 2100 [Wang et al., 2006], which advances the transparency of this important modeling step.
 In addition, PFTs represent important ecological distinctions that are often not presented in LUCC maps, in particular C3/C4 photosynthetic pathways, which are important to hydrologic and carbon cycle studies. The difference in stomatal conductance in C3 and C4 plants produces large differences in transpiration and water use efficiency, and a C4 plant canopy will typically partition more net radiation to sensible heat than latent heat compared to a C3 canopy operating under identical conditions [Still et al., 2003].
 The recent development of continuous fields of vegetation [DeFries et al., 1999; Hansen et al., 2003] overcomes the above problems of translation between land cover types and PFT classifications, in that parameters required for PFT classes (e.g., growth form, leaf type, leaf duration) are measured directly from satellites, depicted by their percent presence by pixel. Data sets generated by this method are of high quality, and some modelers such as Bonan et al.  have adapted their models to use these percent fields, providing, for example, improvements in LSM modeling of the surface climate [Lawrence and Chase, 2007]. While our method has higher information loss in increased error because of the translation of land cover maps to PFTs, the advantage of our approach is that it enables us to examine anthropogenic impact directly.
 The problems with the state of knowledge identified above are in part due to the lack of availability of suitable maps. With the aim to address these problems, we present an assembly of LUCC maps (here, a raster covering the Earth that has a data attribution for all of the land cells) and layers (here, a raster covering the Earth that has “no data” for some of the land cells), tailored to LSM studies, which represent global-scale change from potential land cover (PLC), the vegetation cover that is expected in present-day climate had no human intervention occurred, to current land cover (CLC), the land cover expected to exist at present. The maps have a resolution suitable for state-of-the-art LSMs and Earth system models. This land cover change data set consists of PLC and CLC maps, the difference being described by ALC layers. It is designed for experiments that enable elucidation of direct effects of land cover change, in contrast with experiments that combine temporally varying drivers of climate and land cover change. It is important to note that the land cover changes in the data set are an idealized picture of the differences between climax vegetation communities under current climatic conditions and an interpretation of a range of current land cover maps. Given the idealized nature of the land cover change, it is expected that there will be local areas of unrealistic land cover change.
 The data set has the following key characteristics: (1) has areas of transformed land that is within the 2 standard deviation range of current literature estimates (areas are conserved in all raster manipulations), (2) represents major ALC changes, including hydrologic alterations, (3) has land classes that have an improved and clear correspondence with PFT classes, and (4) has a relatively high spatial resolution in order that the grids can define detailed features, such as small urban areas, reservoirs, or wetlands, and that upscaling to the most state-of-the-art resolution in land surface modeling is possible. The base resolution of the maps is 5 min, with dominant cover classifications in the CLC and PLC maps, and percent cover for the ALC layers. We show how the database can be adapted to PFTs in an LSM, using the case of ORCHIDEE [de Rosnay et al., 2003; Ducoudré et al., 1993a], but it can be used for all the LSMs that have similar or simpler PFT categories. In order to be able to evaluate anthropogenic impact on PFTs, our method links PFTs to land cover so that PFT distribution associated with ALCs and PLCs may be distinguished; a limitation to this method is that it negates the major motivation of using PFTs in describing biome composition: to describe landscape heterogeneity continuously without the constraint of arbitrary defined land cover classes.
2. Assembly of Land Cover Data and Layer/Map Creation
2.1. PLC Mapping
 For the PLC map (Figure 1a), we adapt a widely used map from the University of Wisconsin [Ramankutty and Foley, 1999], generated from the IGBP Discover Seasonal Land Cover Regional data set (SLCR 1.2) which provides detailed classifications of land cover in 1992 for each of five continental regions. Grid cells with more than 50% crop cover or less than 20% dominant potential vegetation were replaced by simulated natural vegetation data as a function of climate and that the vegetation cover is in equilibrium with the present climate for the mid-1990s using the BIOME3 model [Haxeltine and Prentice, 1996], producing a map of dominant potential land cover at a 5-min resolution. The map has 15 land classes (Table 1a), which we have modified to 22 potential land classes corresponding more closely with typical PFT types (Table 1b). The resulting areas for the 22 PLC are given in Table 1b. These modifications were conducted in a Geographic Information System (GIS) and are as follows:
[FAO, 2005; Leff et al., 2004; Lehner and Döll, 2004; Olson and Dinerstein, 1998; Ramankutty and Foley, 1999, 1998; Siebert et al., 2005b; Winslow et al., 2003; N. Ramankutty, 2004; B. Miteva, 2004; A. Strahler, personal communication, 2002]. Grazing land is censored to not occur in barren land, desert, tundra, or shrubland. The classes of C3/C4 composition were chosen on the basis of the distribution of histogram, which had two peaks at 0 and 100%, and a remaining even distribution between 10 and 90%. The classes of 0% covers the peak in the 0–10%, the 30% category covers the 10–50% range (mean of 30%), the 70% category covers the 50–90% range (mean of 70%), and the 100% category covers the second peak in the 90–100% range.
Difference between censored and noncensored grazing land
Wetland loss layer
this study; 9
Irrigated areas layer
 1. Division of savanna into temperate and tropical savanna. In the original map, savanna is a single land cover class that encompasses two types of tree PFTs: tropical broad-leaved raingreen in tropical savannas [Hoffmann and Jackson, 2000], and temperate broad-leaved summergreen in temperate savannas (Wisconsin Botanical Information System, 2006, Habitat Descriptions, University of Wisconsin, Madison, available at http://www.botany.wisc.edu/wisflora/curtis.asp). We thus divided the savanna class into temperate and tropical savannas, using 30 degrees latitude as a limit between the two classes.
 2. Division of mixed forest into temperate mixed forest and boreal mixed forest. The “mixed forest” biome in the original PLC map encompasses two PFT forest categories: temperate and boreal mixed forest. We separate the mixed forests into temperate and boreal mixed forests, using the boundary of “boreal evergreen forest” from other maps [FAO, 2005; Olson and Dinerstein, 1998; Ramankutty and Foley, 1999].
 3. Division of savannas and grasslands based on C3 and C4 composition. As C4 plants are largely confined to the herbaceous growth form [Still et al., 2003], we subdivide the land covers with dominant herbaceous composition, grasslands, tropical savanna, and temperate savanna into C3/C4 classes, using existing maps on the distribution of C3 and C4 grasses for the current climate throughout the globe, as predicted from the SAW algorithm which predicts C3/C4 distribution on the basis of the seasonal timing of water availability with respect to the different C3 and C4 growing seasons [Winslow et al., 2003].
 4. Greenland, Spitzbergen, and Antarctica have been added to the PLC map from the GLCC EDC USGS Olson Classification Map [Loveland et al., 2000] so that the major landmasses are included. These areas contain polar desert/rock/ice and tundra land classes.
2.2. Current Extent of ALC
 We assemble and modify four anthropogenic land cover data sets to produce a layer of the current extent of each ALC. We define major ALC types as those that cover over 1% of the Earth's land surface, as estimated from the literature, and that represent a relatively permanent state of land cover change. These are cropland (both irrigated and nonirrigated), built-up areas, and grazing lands (Table 1b). Anthropogenically burned areas (such as fuelwood burning) were not included because the land cover change in these cases is transient, involving trajectories of successional stages [Batistella and Moran, 2005], or these areas are replaced by croplands or grazing, and then there would be double counting anthropogenic land covers. A layer of tree plantations, industrial forestry, and selective logging [Asner et al., 2005] was not included because a global database of these land use/cover areas was not yet available. Also, the mapped land changes here do not include anthropogenic land cover changes thought to occur from anthropogenically caused climate change, for example, increase in shrubs in the Arctic [Chapin et al., 2005]. A synthesis of areal estimates and mapping products available for ALCs are provided in Appendix A.
 The data sets of ALC were chosen for our study on the basis of the following criteria: that they were produced no earlier than 1990, with preference given to layers with land cover classifications compatible with the PFTs, and to those with a minimum 5-min (∼10-km) grid resolution (Table 1b), where available. The data set is representative of approximately the middle to late 1990s.
 The map of the present-day distribution of global croplands was obtained from the University of Wisconsin [Ramankutty and Foley, 1998], and represents the cropland cover of the world on a continuous scale, depicting the percentage of land in cultivation during the early 1990s for each 5-min grid cell. This data set was developed to understand the consequences of historical changes in land use and land cover for ecosystem goods and services, and was created by synthesizing remotely sensed land cover data with contemporary land inventory data. Another reason that we chose the Ramankutty and Foley  data set to represent croplands was because it contains a new layer of global crop types (e.g., beans, rice, and corn) [Leff et al., 2004] having a distinct C3/C4 composition, and thus which we used to separate cropland into C3/C4 classes, classifying “others” crops as C3, because they consist of mainly C3 plants in the description of the category [Leff et al., 2004].
 The map of the present-day distribution of global built-up land was obtained from the University of Wisconsin (B. Miteva, 2004, Map of Built-Up Areas of the World, SAGE, Madison, Wisconsin, available at http://www.sage.wisc.edu/atlas/maps.php?datasetid=18&includerelatedlinks=1&dataset=18). This map combines the two main sources of the extent of urban and built-up areas: the DMSP/OLS Nighttime Lights [Imhoff et al., 1997] and the IGBP land cover characterization data sets (A. Strahler, Consistent Year Product, Boston University, personal communication, 2002) gridded to percent cover at a 5-min resolution.
 For the map of the present-day distribution of global grazing lands we used a map from the University of Wisconsin (N. Ramankutty, Map of Grazing Lands of the World, 2004). The data set is a 30-min grid that has been converted to a 5-min grid, with percent cell occupied by grazing land. This data set was selected also because it is complementary in extent to the map of croplands, with few overlapping cells. We modified the grazing layer so that it is divided into C3/C4 categories, using the same method as used for natural grassland in the potential vegetation map.
 In addition to its typical grassland setting, grazing occurs also in harsher environments such as tundra, shrubland, and desert. At present, many models parameterize grazing land as a form of grassland. Therefore, if areas such as open shrubland were included as grazing land in such model studies, the model would simulate grazing of open shrubland as a conversion to grassland, and spurious ecological responses might result, such as an erroneous increase of evapotranspiration. To address this problem, we have created a “censored” version of grazing land that does not allow grazing in areas with lower productivity or frozen soils. Thus, we have two grazing scenarios: censored, where grazing on shrubland, desert, and tundra is removed, for use in most LSMs, and noncensored, where grazing on shrubland, desert, and tundra is included, which we include when we want a global survey of anthropogenic impact. It is known that grazing indeed does occur in these less productive regions; for example, since 1978 in the Tibetan Plateau that contains tundra, sheep and cattle production have increased by 106 and 249%, respectively [Cui et al., 2006; Du et al., 2004].
2.3. Current Land Cover Map
2.3.1. Map of Dominant Current Land Cover
 We created a map of CLC (Figure 1b) through the combination of ALC layers and the PLC map. In order to make a single CLC map, as required by many LSMs, in lieu of having a separate layer for each ALC, we integrated three different ALC layers into one layer (cropland, censored grazing land, and built-up land). We did not include the inundated land layer because this CLC map is designed for LSMs like ORCHIDEE which do not contain an open water PFT. ALC layers needed to be converted from the continuous percent cover to dominant land cover at the 5-min resolution in order to create a single CLC map. A threshold percentage is used to convert percent cover to presence/absence, chosen so that the original area of the ALC is globally conserved (39, 19, and 29% for cropland, built-up land, and censored grazing land, respectively). In creating the dominant land cover map, if more than one of built-up, censored grazing, or cropland occupy the same cell, priority of cell occupation is given first to built-up land, then to cropland, then finally to censored grazing land, chosen to represent increasing total area. Finally, to create the CLC map, we replaced 5-min cells in the PLC map with 5-min dominant cover ALC cells. The ALC cells form 33% of the CLC land surface (Table 2).
Table 2. Comparison of Anthropogenic Land Cover Areas in This Database and Other Widely Used Land Cover Mapsa
 Because of its potential importance to the Earth system, we have generated a layer of wetland loss (Figure 1c). To do so, we modified a map of global wetland distribution from University of Kassel (CESR) Global Lakes and Wetlands Database (WELAREM1) [Lehner and Döll, 2004]. WELAREM1 is a global 1-min (∼1.8-km) map in dominant cover of wetlands, lakes, and reservoirs and was derived by combining various digital maps (Environmental Systems Research Institute (ESRI), 1993 (Digital Chart of the World 1:1 Mio., Redlands, California, data obtained on four CDs, also available at http://www.maproom.psu.edu/dcw/): wetlands, lakes, and reservoirs; ESRI, 1992 (ArcWorld 1:3 Mio. Continental Coverage, Redlands, California, data obtained on CD): wetlands, lakes, reservoirs, and rivers; World Conservation Monitoring Centre, 1999 (Digital Wetlands Data Set, Cambridge, United Kingdom; see http://www.geo.uni-frankfurt.de/ipg/ag/dl/forschung/Global_Water_Modeling/WELAREM1/index.html): lakes and wetlands; Vörösmarty et al. : reservoirs; and attribute data (International Commission on Large Dams (ICOLD)  and Birkett and Mason : lakes and reservoirs)). To obtain the map of wetland loss, we removed the dominant cover 1-min wetlands that exist in the same 5-min cell with the ALCs (cropland, censored grazing land, and built-up land). The cells that were removed (wetland loss) are expressed as a percentage cover for a 5-min cell (Figure 1c). We used the censored grazing land so that wetlands would not be removed by grazing in shrubland, deserts, and tundra, because in these dry areas it is unlikely that wetlands in these areas will be drained for grazing purposes; instead the wetlands may be used as oases. The bulk of such areas is located in central Australia, the Tibetan Plateau, and in the Arabian Peninsula. We treat the wetland loss as separate from the PLC/CLC map because wetlands coexist with other land covers; for example, wetlands may occur in grasslands, forests, or tundra.
 For the map of present-day distribution of anthropogenic inundated land/reservoirs, we used a map from University of Kassel (CESR) Global Lakes and Wetlands Database (WELAREM1), as described in section 2.3.2 [Lehner and Döll, 2004]. Inundated land in reservoirs is treated as a separate hydrologic layer because some LSMs, like ORCHIDEE, do not have parameterizations for this land cover type.
 Finally, in our analysis we include a layer of irrigation, based on an existing map produced by Siebert et al. [2005a]. The layer documents amount of area equipped for irrigation in the 1990s as a percentage of the total area on a raster with a resolution of 5 min. Percent cover was converted to dominant land cover, using a threshold of 5.0%, chosen to preserve the estimated area of irrigation at approximately 17% of all cropland [Wood et al., 2000].
3. Analysis and Validation of Land Cover Transformations
3.1. Analysis of Land Cover Transformations
 First, by comparing PLC (Figure 1a) with CLC (Figure 1b), we obtain an estimate of the spatial distribution of anthropogenic alteration of the Earth's surface. We see that most of the productive temperate and tropical areas have been converted to human use, except for mountainous or very dry terrain and large tropical forest areas around the Zaire and Amazon watersheds, and parts of Southeast Asia and the Northern Hemisphere evergreen forests. The amount of forest lost is likely underestimated in these areas because, as described above, we do not include industrial forestry or conversion to tree plantations. For example, in Myanmar it is known that large parts of the forest have been lost to deforestation and shifting cultivation [Brunner et al., 1998], and in western Canada it is known that extensive areas have been deforested for industrial forestry.
 Next, by employing a GIS analysis of the LUCC data set (PLC and CLC maps and ALC layers), we develop a perspective of the nature of the current extent of global LUCC. This reveals that some natural biomes have been almost completely converted to anthropogenic use (Table 1b). Grassland biomes are the most impacted in terms of percent loss: almost all (92%) of the original natural grassland area has been converted. Approximately 75% of natural savanna has been transformed to anthropogenic uses; temperate savanna with mixed C3/C4 grasses (30% C3 grass) is almost completely lost. Forests also have also undergone a large percent loss, ranging up to 68.3% for tropical deciduous forest and woodlands. In terms of total area of change, anthropogenic activities have caused the greatest losses of the natural states of tropical savannas and C3 grasslands (Figure 2a). In contrast, the land covers with the biggest areal increase are C4 grazing land, followed by C3 cropland (Figure 2a).
 Also, the mapping database provides insight into which potential biomes each ALC has replaced (Figure 2b) Over half of cropland occurs in previously forested areas (predominantly in Europe, China, and India), and a large proportion of cropland occurs in previous mixed C3/C4 grassland (North America and Russia). As consistent with most built-up land occurring in developed countries, we find that the majority of built-up land occurs in the (potential) temperate forest biomes. Grazing land predominantly occurs in the (potential) temperate grasslands, open shrublands (when considering grazing in less productive areas), tropical savannas, and tropical grasslands. For wetland loss, the data indicate that almost a quarter has occurred in the tropical savanna biome, in South America and in sub-Sahelian Africa, and another quarter in the grasslands of the Northern Hemisphere (in particular in the North American prairies). Irrigation occurs most often in lands that were converted from tropical deciduous forest (India), temperate mixed forest (China), and shrublands (central/western Asia, west/central North America). Most of the reservoirs occur in the boreal evergreen forest region in Russia and Canada, associated with major hydropower projects in the northern rivers.
 Finally, the assembled maps indicate the associated anthropogenic causes for the losses in each PLC (Figure 2c), with the caveat that forest converted to abandoned land or tree plantations is not considered here. We see that savannas and tropical forests have very similar patterns of ALC replacement. Conversion to grazing land is the principal driver for loss of natural grasslands (82% from grazing) and for tropical savanna (92% from grazing); however, conversion to cropland is the major cause of loss of natural temperate savanna (64% from cropland). Loss of temperate and boreal forests is due mostly to cropland conversion (∼80% of converted land), where deforestation of tropical forests tend to result almost equally from conversion to cropland and grazing land. Boreal forest is the only natural biome that has a notable fraction lost to inundated lands. The total loss of potential forest cover for all biomes is estimated to be 1.4 E + 13 m2, or ∼23% of the original forest cover, slightly less than the FAO estimate of original forest cover loss of 3.95 E + 13 m2 [FAO, 2005]; the difference may be explained because our estimate does not take into account forest loss from tree plantations. However, our estimate of total forest loss 1.43 E + 13 is slightly larger than estimates of total forest loss in the past 300 years (∼0.7 to 1.1 E + 13 m2) [Foley et al., 2005].
 Natural shrubland has been estimated to be largely replaced by anthropogenic land cover, estimated to currently be 10.5% of original (i.e., potential) shrubland area [Asner et al., 2004]; if the “noncensored” grazing was included in our estimate of shrubland loss, shrubland we find that, like Asner et al., most of the open shrubland has been converted to grazing land (97%). Our observation that grazing land predominantly occurs in the temperate grasslands, open shrublands (when considering grazing in less productive areas), tropical savannas, and tropical grasslands is consistent with previous estimates [Asner et al., 2004].
3.2. Validation of Land Cover Maps and Layers
 As a validation we compare the areas of land cover and of land cover change presented in this database with estimates from the literature. First we estimate the mean and variance global area of ALC as a percent of the Earth's land surface, abbreviated here as %ALC (not including permanent ice areas). By summing the mean areas of cropland, grazing land, built-up land, burned land, tree plantations, and inundated land as gathered from the literature (Figure 3 and Appendix A), we estimate mean %ALC to be 44%. This value falls within the 39–50% range previously estimated [Vitousek et al., 1997], and it is consistent with an estimate that wilderness covers 44% of the globe [Mittermeier et al., 2003], in that the remaining ∼12% of land may be accounted by fragmented land or land near roads that was not included in either occupied or wilderness classification. The uncertainty in the global estimate of %ALC is large, ±17% at a 95% confidence interval, reflecting the high degree of uncertainty in all area parameters except those related to tree plantations (a parameter whose low uncertainty likely reflects an absence of independent data sources). The confidence interval was estimated using stochastic Monte Carlo simulations in which each parameter was allowed to vary randomly constrained by its mean and estimated variance (Appendix A). The estimate of variability in our knowledge of %ALC was estimated by repeating these calculations 1 million times.
 As compared to literature means, the %ALC in the CLC map is 41% (Table 2) (uncensored grazing), very close to the literature mean of 44%. Our CLC map value is slightly less than the literature mean, which is explained by the absence of inundated land, burned areas, and tree plantations in the CLC map (the total of literature means of cropland, built-up land, and grazing land is 39% of the Earth's land surface). Note that the areal estimates for each ALC in this database are included in the calculation of the literature averages (Appendix A).
 In this study, total cropland area is 1.8 E + 13 m2, close to the literature average of 1.7 E + 13 m2. The area of uncensored grazing land in this study is 3.6 E + 13 m2, slightly more than the literature average. There is a wide range of estimates of built-up land in the literature; the area of built-up land in this study (5.0 E + 11 m2) lies in the middle of the literature estimates. The area of inundated land is less than the one estimate found in the literature (Appendix A).
 There is a wide variation in %ALC in other commonly used land cover databases (Table 2). We find the HYDE database [Klein Goldewijk, 2001] has a total anthropogenic area (36%, Table 2) that is similar, if just slightly less, to the literature estimate of %ALC, and thus to our study. Our estimate of %ALC is slightly larger than the HYDE database, mostly deriving from the cropland being larger in extent. Other commonly used land cover databases, however [Houghton et al., 1983; Loveland et al., 2000; Matthews, 1983; Richards, 1990], have estimations of anthropogenic extent below the 2 standard deviation of the literature mean (Table 2). Two of the data sets are at least a decade older, which may explain why the cropland extent is slightly lower than the literature mean; yet it is the absence of accounting for grazing land in these data sets that explains why they are much lower than the literature average of total anthropogenic land cover.
4. Conversion of Land Covers to PFTs
 As they were designed, the PLC and CLC map land cover classes are easily converted into PFT categories. And the description of land cover change in terms of change in PFTs provides insight into the nature of change in ecological functioning caused by LUCC. As an example of PFT conversion from this LUCC data set (Table 3), we use a typical PFT scheme from the IPSL ORCHIDEE LSM [de Rosnay and Polcher, 1998; Ducoudré et al., 1993b; Verant et al., 2004]. Even though the translation of land cover classes to PFT classes is relatively smooth, there are still assumptions that need to be made in the conversion, in particular, the assignment of percent bare soil to each land cover type (Table 3). Note that parameterizations of built-up land are still very crude in LSMs, and are usually parameterized as bare soil, as is the case here. Since grazing land was censored for the LSM simulation, the PFT areas reflect the censored version of grazing land.
Table 3. Conversion of Vegetation Land Cover Classes to Plant Functional Types Used in the LMD Land Surface Model ORCHIDEE, With Reference to the Literaturea
PFT and land cover classes are the same. The PFT fraction falls within the limits for these land covers given by Crucifix et al. .
The PFT forest composition was estimated from the Atlas of Russia's Intact Forest Landscapes (available at http://www.forest.ru/eng/publications/intact). The boreal deciduous forest class is mainly in eastern Siberia, in the larch forests.
Assume 50% coniferous and 50% deciduous, on the basis of previous ORCHIDEE classification. The PFT fraction falls within the limits for these land covers given by Crucifix et al. .
PFT and land cover classes are the same, percent bare soil taken from original ORCHIDEE scheme, as it is within the range of grasslands given by F. B. Fisher et al. (1998, Montana Land Cover Atlas, unpublished report, viii + 50 pp., Montana Cooperative Wildlife Research Unit, University of Montana, Missoula). The PFT fraction falls within the limits for these land covers given by Crucifix et al. .
The shrubland biome is dominated by shrubs with small but thick evergreen leaves; therefore we choose broadleaf evergreen as the tree species. Since shrublands occur in parts of South America, Western Australia, central Chile, and around the Mediterranean Sea, we chose the type of tree to be “temperate.” We choose broadleaf evergreen for the “tree” (Community and Ecosystem Dynamics, available at http://www.estrellamountain.edu/faculty/farabee/biobk/BioBookcommecosys.html. ORCHIDEE does not have a shrub PFT. Grass must be added so that shrubland does not behave like a forest, even though there is very little grass in the actual biome. The percentage grassland is based upon Ngo-Duc et al. . C3 grass is chosen, because shrubs they are intended to simulate are C3.
 The resulting inventory of changes in PFTs (Table 4 and Figure 4) provides a different picture than that from land cover change. Tropical broadleaf raingreen forest has the largest losses associated with anthropogenic land cover change, and human activities have caused the loss of about 72% of this plant functional type. Temperate broadleaf summergreen forest undergoes the next largest losses of all PFTs, with half of this PFT being removed by human activities. The total loss of trees PFTs is 1.9 E + 13 m2, over a quarter of the potential amount of trees on the planet. This area of lost trees is larger than estimates of forest loss in the land cover change inventory because it includes additional trees lost in savannas. The boreal forest PFT has undergone a very small reduction compared to other forest types (3.5% as compared with tropical forests of 42% loss, and temperate forests of 48% loss); the losses of this PFT are likely underestimated because of the lack of tree plantation mapping.
Table 4. Changes in Global PFT Distribution From Potential to Current Land Cover Scenariosa
Area (1012 m2)
Percent of Total
Area (1012 m2)
Percent of Total
An example using ORCHIDEE PFTs. Grazing land is censored.
Tropical broadleaf evergreen forest
Tropical broadleaf raingreen forest
Temperate needleleaf evergreen forest
Temperate broadleaf evergreen forest
Temperate broadleaf summergreen forest
Boreal needleleaf evergreen forest
Boreal broadleaf summergreen forest
Boreal needleleaf summergreen forest
 Grassland PFTs undergo a much smaller change than do forests from anthropogenic land cover (−6.1%), as some natural grasslands and savannas are replaced by grazing land, which includes the grassland PFT. The global ratio of C3/C4 plants (more water intensive versus more water efficient) undergoes an 8% increase from 1.8 to 1.9 in the potential vegetation scenario, suggesting that the total water use efficiency of crops and grasses has slightly decreased.
 Of all PFTs, human activity causes the largest increase in C3 cropland (Figure 4). Next largest increase is bare soil, with an increase of 19%, reflecting the increased land disturbance in anthropogenic land covers. C4 grassland also experiences an increase in area, related to replacement of C3 forests with C4 pasture grasses in the tropics (as given by Still et al. ).
 The PFT classes which undergo the largest losses from anthropogenic activities are not the same as the land cover classes which undergo the largest losses, which were found to be grasslands and savannas (Figure 2a). The reason for the differences in results from when one considers land cover classes and PFTs is that, in some cases, the ALC that replaces the PLC has the same dominant PFT, as is in the case with tropical savanna (C4 grass) being replaced by grazing (C4 grass), although there is concurrent increase in bare soil and reduction in tree cover. Note that the functional changes caused by grazing on tundra, open shrubland, and desert land are not included in this inventory, but would likely result in an increase in percent bare soils for these biomes, yet some nutrient-limited ecosystems may see a reduction in bare soil, as grazing may add nutrients to the system.
5. Discussion and Conclusions
 The results are limited by the quality of the maps used, including the potential vegetation map, and each ALC map. Further, generalization of land cover at a global scale inevitably requires some assumptions that will not necessarily hold at finer resolutions, and this database is not designed for use at finer scales. Land cover data sets are inherently uncertain, both in mapping the correct location of land cover [Iwao et al., 2006] and the total area of land cover altered (Figure 3).
 This work highlights the need for improvements on land cover mapping. Grazing land has a huge impact upon the Earth, and improvements in mapping of grazing areas are important. And it is important to improve how LSMs account for grazing, for example, by taking into account grazing intensity. Further, we also underline the importance of the creation of a map of global tree plantations and “industrial/managed forests.” The current lack of available maps on global tree plantations is unfortunate, as it has been estimated that timber plantations in North America and oil-palm plantations in Southeast Asia now cover 1.9 E + 12 m2 worldwide [Foley et al., 2005] and should continue to increase, as signatory countries work to meet carbon reduction requirements stipulated in the Kyoto Protocol, and as large tracts of rain forest and wetlands are being replaced by palm tree plantations in areas such as Malaysia and Brazil to meet the rising demand for biofuel oil [Rosenthal, 2007]. And improved mapping of boreal forest loss through industrial forestry is also important because boreal forest loss has been identified to have important impacts on surface temperature because of altered albedo [e.g., Betts et al., 2007].
 One important step in modeling land cover in LSMs is the choice percentage bare soil assigned to occur in each land cover, as required in the reclassification matrix between land covers and PFTs. The percentage of bare soil drives the major land surface fluxes and properties, determining, among others, LAI and albedo. However, in general, the percent bare soil that is included in each land class is not presented. Recently, Wang et al.  have made this bare soil consideration transparent in their land cover/PFT data set for modeling studies by publishing the conversion table between land cover classes and PFTs; however, in this table, they do not discuss how the percent bare soil was chosen. For our reclassification matrix, we chose the bare soil percentages on the basis of literature estimates (Table 3), albeit regional in nature. With our simulation, we see percentage bare soil increase with human impact (Table 4) in line with expectations that human disturbance to land cover causes an increase in bare soil, even though our bare soil percentages for cropland and grazing land are conservative (Table 3). In contrast, the land cover change data set presented by Wang et al.  shows a decrease in bare soil with human impact, which may be due to their noninclusion of grazing land in the study, and the percent bare soils being relatively high in natural grassland and savanna and quite low in cropland. Indeed, the main difference between our classification of grazing and grassland is the percentage bare soil. Such decisions can have major impacts on the simulation results and need to be transparent. A global standard of percent bare soils for land cover classes is needed, and may help to reduce the large variability in LSM output that are driven by the same forcing data, as found in the GWSP2 program [Dirmeyer et al., 2006]. A good place to start would be to generate statistics on bare soil for major land covers from the data on global fractional cover of bare soil that is independently described from MODIS and AVHRR satellite imagery [DeFries et al., 1999; Hansen et al., 2003].
 Here we have created a mapping LUCC database that provides a coherent comparison of potential and current land cover, for which all change is limited to anthropogenic action, thus allowing for direct tests of impacts of LUCC. Here we have improved the state of the art by making a mapping data set available to modelers that increases awareness of the importance of area and in the inclusion of a broad range of hydrologic major anthropogenic land cover changes, including significant alterations to the African savannas, which are not seen in many data sets [e.g., Betts et al., 2007], and which can have large impacts on the water cycle [Sterling, 2005]. The relatively fine resolution (5 min) of the map allows it to be used in a variety of global LSMs.
 The database is tailored for use in LSMs to assess the direct impact of LUCC on land surface fluxes, water and carbon cycle, and climate, for both coupled and uncoupled simulations. It has a transparent conversion to PFTs that enables improved interpretation and comparison of LSM studies on LUCC. The straightforward conversion of land covers to PFT categories is illustrated for the ORHCHIDEE LSM to assess the impact of LUCC on PFT distribution.
 Our land cover change database documents the large impacts humans have had on the land surface of the planet, reducing some biomes by up to 92%. It also provides insight into how humans have changed the ecological functioning, such as by reducing the natural tree cover by 27%, and increasing the proportion of C3 to C4 plants. Thus, the biome with the greatest losses from the natural state are the natural grassland biomes, yet the largest changes to ecological functioning in the world are resulting from loss of forests, because grazing does not change the dominant functional type of grassland, although bare soil is increased.
 We emphasize that an important step to address land cover change uncertainty in LSM studies is to document fully anthropogenic land cover areas (%ALC) that are used, to aid in the interpretation and intercomparison of results. We show here that the LUCC database presented here reflects well our current understanding of the extent of human transformation of the Earth surface, yet there is a wide range of %ALC in other commonly used land cover databases, all of which underestimate the %ALC, some by more than 2 standard deviations of the estimated literature mean of 44%.
Appendix A:: Literature Estimates of Area of Anthropogenic Land Covers
 We assemble estimates of transformed areas for each anthropogenic land cover type, including the following human-dominated land covers in our analysis: agriculture, tree plantations, built-up land areas, grazing land, land inundated by dams, and biomass burning, on the basis of the criteria that they each cover at least 0.1 percent of the Earth's terrestrial surface and reflect direct use (Tables A1–A5). Area estimates are constrained to be recent (published since 1990). We do not include “indirect” anthropogenic cover changes such as desertification, inadvertent draining of the Aral Sea, fire suppression, climate change, pollution induced changes, deforested lands that are abandoned, or lands impacted by war. Our estimates of land cover change are thus conservative. In the estimate of the area of grazing land the four estimates do not appear independent.
C. M. Biradar et al., A global map of rainfed cropland areas (GMRCA) using time series data from multiple satellite sensors, submitted to International Journal of Applied Earth Observation and Geoinformation, 2008, in review.
1.53 e + 13 m2
derived from work on GIAM, in review, and C. M. Biradar, personal communication, 14 April 2008
represents the area in croplands during the early 1990s for each grid cell on a global satellite 5-min resolution latitude-longitude grid; combined a satellite-derived land cover data set with a variety of national and subnational agricultural inventory data; uses adjusted FAO data from Alexandratos ; does not include plantations and shifting cultivation
Table A2. Global Built-Up Areas (or Human-Occupied Land)a
Maps of global built-up area include [Elvidge et al., 2001; Ataman et al., 2006; B. Miteva, 2004; A. Schneider and M. A. Friedl, A global map of urban areas from MODIS data, manuscript in preparation, 2008].
sum of area of tropical and temperate forest plantations
1.2e + 12 m2
2.0e + 11 m2
 We thank M. Mancip, T. d'Orgeval, and N. Ramankutty for helpful comments. We thank J. Chen for research support. This work was partly funded by a Chateaubriand Fellowship for Scientific Research from the Office for Science and Technology of the Embassy of France in the United States and a Marie Curie Intra European Fellowship 09949.